Artificial intelligence enabled spectral-spatial feature extraction techniques for land use and land cover classification using hyperspectral images – An inclusive review

The growth of artificial intelligence techniques such as machine learning and deep learning facilitates the hyperspectral image processing applicable in developing various remote sensing applications such as Change detection in Land Use and Land Cover (LULC) classification, Evaluation of the nutritional content, and health of the crops in Agriculture. However, Hyperspectral imaging is frequently utilized in remote sensing and earth observation applications to identify environmental changes. One of the key tasks in hyperspectral image classification is feature extraction. This paper gives a comprehensive review of the recent hyperspectral image feature extraction techniques for LULC. This study aims to identify the open issues, research challenges, and future directions that will help researchers develop efficient feature extraction techniques for better LULC hyperspectral image classification. The performance of the state-of-the-art feature extraction techniques for hyperspectral images is analyzed in terms of the overall accuracy, average accuracy, and kappa coefficient across the benchmark datasets, namely Indian Pines, Pavia dataset, and Salinas dataset. From the analysis, we observe that in all the benchmark datasets, the framework 2D + 3D CNN with spectral-spatial integration not only extracts the comprehensive features but also increases the classification accuracy with less computational complexity compared to other competing frameworks. Both 2D CNNs and 3D CNNs are utilized for extracting features and patterns from data with multiple spectral bands, and each architecture has its advantages and challenges. 2D CNNs are more common and computationally efficient, while 3D CNNs capture spatial-spectral correlations more directly.